Regression detection
Most accessibility scanners report flat counts. The regression detection module reports change: which violations are new since the last deploy, which deploy introduced them, which AI or human author wrote the offending code, and which root-cause cluster they belong to. This is a Detect-tier module inside the agentic compliance architect — the diff layer of the eleven scanner-axes.
What it does
-
Cross-deploy diff engine
Compares each scan against the prior baseline by stable finding-fingerprint — murmurhash3 of rule id + role + accessible name + nearest stable selector. Survives CSS class hashing and framework wrapper injection.
-
Root-cause clustering
47 raw violations collapse into 4 clusters: "missing
aria-labelon icon buttons insrc/components/Form/", "low-contrast warning text in design-system token--warn-fg", and so on. Engineering reads root causes, not row-level counts. -
Per-component trend
Trends each component over time: was
<FormField/>regressing for three deploys before the failure? Was last week's spike driven by one merge? This answers that without manual triage.
Layer mapping
| Axis / Layer | Role | Direction |
|---|---|---|
| L3 Diff layer | Canonical home of regression detection — deploy-over-deploy violation diff + cluster. | — |
| Multi-domain scan | Source of ScanEvent input. Regression detection consumes successive scan snapshots. | Consumes |
| AI authorship attribution | Joins each new finding to the AI or human author that introduced the regression. | Consumes |
| LLM remediation cascade | Receives root-cause clusters; remediation targets the cluster, not row-by-row. | Feeds |
| CI/CD gate | Reads the new-since-baseline subset to apply differential AI-vs-human thresholds. | Feeds |
Why this is different
Most accessibility tooling answers "how many violations does my
site have?" Regression detection answers "which 31 of those 47
violations are new since deploy SHA xyz, clustered
into 4 root causes, all introduced by Copilot in
src/components/Form/*." That is the difference between
a number on a dashboard and an actionable remediation queue feeding
the LLM cascade.
Filed IP
ARIADA holds filed-IP positions covering the cross-deploy regression detection methodology — including cluster-root-cause analysis, per-component trend attribution, and deploy-SHA-linked violation tracking — underlying this module. Provisional application only; conversion to non-provisional pending.
Application numbers, claim counts, and PCT deadlines are available for accredited-investor due diligence on the Legal & IP page.
Cross-references
- Detect-tier siblings: AI artifact audit, scan visualization.
- Upstream input: Multi-domain scan.
- Author attribution: AI authorship attribution.
- Downstream consumers: LLM remediation cascade, CI/CD gate.
- Architecture overview: Technology architecture.
Source-level remediation only — agentic suggestions are not autonomous deployments; pull requests require client merge. Not a legal certification body. For accredited certification, customers consult a notified body or registered auditor.